Detecting abandoned citrus crops using Sentinel-2 time series. A case study in the Comunitat Valenciana region (Spain)
Derechos de accesoopenAccess
MetadataShow full item record
Cita bibliográficaMorell-Monzó, S., Sebastiá-Frasquet, M. T., Estornell, J. & Moltó, E. (2023). Detecting abandoned citrus crops using Sentinel-2 time series. A case study in the Comunitat Valenciana region (Spain). ISPRS Journal of Photogrammetry and Remote Sensing, 201, 54-66.
Agricultural land abandonment (ALA) is becoming a growing phenomenon around the world that needs to be monitored and quantified. A massive abandonment of citrus orchards has been experienced in the last years in the Comunitat Valenciana (CV) region (Spain) driven by different socio-economic factors. Therefore, developing time and cost-efficient methods for monitoring ALA is a priority. Citrus are a perennial crop trees which make orchards have low spectral variation during the year. In the CV region, they are planted in relatively small parcels, thus creating a highly fragmented and heterogeneous landscape. This study proposes a machine learning-based classification framework that uses annual time series of spectral indices extracted from Sentinel-2 images to identify crop status at parcel level. The method is based on features extracted from the reconstructed OSAVI and NDMI time series used to train a Random Forest classifier. Then, a parcel-based classification is performed using the parcel boundaries and the probabilities of belonging to each category for the full pixels inside the boundaries. The research assessed the potential to identify three statuses of crops (non-productive, productive, and abandoned). Results on three different temporal and spatial datasets provided an overall accuracy ranging from 89 to 92 %, demonstrating the importance of multi-temporal data to identify the abandonment of perennial crops. Furthermore, we studied the ability of the model to be spatially and temporally transferred. Limitations to recall the abandoned parcels when using models trained in other areas or time periods are exposed, opening the way to model improvements.